An architecture for planning, inverse planning, and inference in planning, using PDDL and Gen.
Initially developed under the MIT Probabilistic Computing Project and the MIT Computational Cognitive Science lab. Now maintained by the Cooperative Systems & Intelligence (CoSI) lab at NUS.
To use this library in your own projects, press ] at the Julia REPL to
enter the package manager, then run:
add PDDL SymbolicPlanners
add Gen GenParticleFilters
add PDDLViz GLMakie
add https://github.com/cosilab/InversePlanning.jl.gitTo explore the examples provided in this repository, clone this repository,
press ] at the Julia REPL to enter the package manager, then run the following
commands:
activate examples
dev ./
instantiateThis will activate the examples directory as the project environment, set up
your cloned copy of InversePlanning.jl as a dependency, and install any
remaining dependencies.
InversePlanning.jl can be used to model agents that perform model-based heuristic search to achieve their goals. Below, we visualize a sampled trace for a replanning agent that interleaves resource-bounded plan search with plan execution:
We can then perform goal inference for these agents:
Notice that the correct goal is eventually inferred, despite backtracking by the agent. This is because we model the agent as boundedly rational: it does not always produce optimal plans. Indeed, this modeling assumption also allows us to infer goals from failed plans:
Because we use the Planning Domain Definition Language (PDDL) as our underlying state representation, our architecture supports a large range of domains, including the classic Blocks World:
For more details about the modeling and inference architecture, consult our paper:
T. Zhi-Xuan, J. L. Mann, T. Silver, J. B. Tenenbaum, and V. K. Mansinghka, “Online Bayesian Goal Inference for Boundedly-Rational Planning Agents,” Advances in Neural Information Processing Systems, Jun. 2020.
Full example code for several domains can be found here: Gridworld; Doors, Keys & Gems; Block Words